Training Status: Track and Report ML Model Training Progress with ETA Estimation

Question

## Problem

Implement a `TrainingStatus` class that tracks an ML training job's progress across epochs. It should record loss and accuracy per epoch, estimate time remaining (ETA) based on elapsed time, and detect if training has stalled (no improvement in loss for `patience` epochs).

```python
import time

class TrainingStatus:
    def __init__(self, total_epochs: int, patience: int = 5): ...
    def record_epoch(self, epoch: int, loss: float, accuracy: float) -> None: ...
    def eta_seconds(self) -> float: ...
    def is_stalled(self) -> bool: ...
    def best_epoch(self) -> int: ...
```

**Example:**
```
ts = TrainingStatus(total_epochs=100, patience=3)
ts.record_epoch(1, loss=0.9, accuracy=0.6)
ts.record_epoch(2, loss=0.8, accuracy=0.65)
ts.record_epoch(3, loss=0.82, accuracy=0.64)
ts.record_epoch(4, loss=0.83, accuracy=0.63)
ts.is_stalled()   -> True   # loss hasn't improved for 3 epochs
ts.best_epoch()   -> 2
```

## Follow-ups

1. How would you compute a smoothed ETA using an exponential moving average of per-epoch durations instead of a simple average?
2. What changes if you want to track multiple metrics (e.g., val_loss vs. train_loss) and stall is defined on val_loss only?
3. How would you serialize the training status to disk so a crashed job can resume?
4. What is the difference between early stopping based on patience and a learning rate schedule with warmup/decay?

Training Status: Track and Report ML Model Training Progress with ETA Estimation

Interview Experience

Full Details

About This Question

Topics

More Vanta Interview Questions